Why Shobdo Never Stores Your Video?
The short version
Most camera AI sends video to someone else's cloud and keeps it there. Shobdo turns video into a written log as it happens and discards the video. The footage stays on the store's recorder. That design protects privacy, and it removes the biggest cost in the system: storing video.
- Eight cameras produce 4 to 8 TB of video a month, and storing it in the cloud costs hundreds of dollars.
- Verkada and Ring both showed what happens when customer video sits on a vendor's servers.
- When the AI needs a clip, it asks the store's recorder; if the clip is gone there, there is no other copy.
- The only video kept on our server is the short clip with an alert, and it expires within 24 hours.
In this post
The video surveillance AI market is projected to reach $49 billion by 2035, growing at over 20% annually. Nearly every video surveillance AI product works the same way: cameras record, footage goes to someone else's cloud, and their AI processes it there. The video sits on their servers, sometimes indefinitely.
We built Shobdo differently. Client video is not stored on our servers. That is not a privacy toggle buried in settings. It is how the system is designed from the ground up.
The problem with storing video in the cloud
A single 1080p camera running 24/7 with H.265 compression generates roughly 30 to 50 GB per day, call it 1 TB per month. A small store with eight cameras produces 4 to 8 TB monthly. At AWS S3 Standard rates of $0.023 per GB, storing 8 TB costs about $184 per month, just for storage. Add egress fees for actually retrieving that footage ($0.09 per GB for the first 10 TB), and reviewing even a fraction of it gets expensive fast. Multiply by months of retention, and the bill becomes the dominant cost of the entire system.
Then there is the question of who else has access to that footage.
In 2021, hackers breached Verkada and accessed 150,000 live camera feeds from hospitals, schools, police departments, and corporate offices, using a single exposed credential. The FTC later required Verkada to pay a $2.95 million penalty. Separately, the FTC charged Ring with failing to restrict employees' and contractors' access to customer video (one employee viewed thousands of recordings from customers' bedrooms and bathrooms) and with using customer videos to train algorithms without consent. Over 55,000 customers experienced account compromises. The $5.8 million settlement followed.
These are the predictable consequence of a design where the video leaves the customer's control.
What Shobdo does instead?
Shobdo's AI watches camera footage and converts it into structured text descriptions: human-readable event logs that record who appeared, what they did, and when. Only those text descriptions are stored on our servers. The raw video stays on the client's local storage.
Industry estimates suggest that less than 1% of surveillance footage is ever actively monitored or reviewed. A human operator's ability to detect incidents drops by 95% after just 20 minutes of watching monitors. The vast majority of recorded video exists solely as insurance: hours of empty aisles and quiet hallways, retained in case something needs to be checked later.
Shobdo replaces that approach. Instead of storing everything and searching later, our AI processes footage as it happens and writes it down as text. A store owner can open the whiskey-aisle camera and read its log for the afternoon. Each line points to the exact clip on the store's own recorder. Those lines can be searched by words and by time today. An AI agent that runs that search for you from a plain question is planned.
When the AI does need to look at a specific video clip (say, the text log mentions someone lingering near the register but the store owner wants to see what they were holding), it requests that clip from the client's local storage on demand. If the clip has been deleted or overwritten, the request simply fails. There is no independent copy. The client's video storage is the single source of truth.
One deployment today, one guarantee
Not every store owner can afford a GPU. Requiring one would price out the businesses we built this for.
So today, video is sent to our server, where the AI turns it into text. The video is processed in memory and discarded. Only the text descriptions are stored in the database. This keeps the price affordable. The store gets the same AI without buying any hardware.
Some stores may want video never to cross the network. Running the AI on a GPU inside the store fits the same design, since only the text would then leave the building.
Either way, video is not stored on our servers. The one exception is the short clip attached to an alert, which sits on our server only until your phone has fetched it and expires within 24 hours.
What this does to the price
The privacy design has a direct economic consequence.
The cloud storage costs described above (hundreds of dollars per month per store, before retrieval or retention) would be the single largest line item in the system. By storing only text descriptions, we eliminate it almost entirely. An event log entry is roughly 0.1 to 2 KB. A month of activity from eight cameras might produce a few hundred megabytes of text. The cost difference is several orders of magnitude.
That saving is what makes Shobdo's pricing work for a store with four cameras and tight margins.
Shobdo is an AI agent for surveillance cameras. Shobdo can send alerts on events you describe using plain English. Learn more or book a conversation.
